Transcript Selling an Idea or a Product

Network Protocols
Protocol: agreement between two parties as to how
information is to be transmitted.
 Example: system calls are the protocol between the
operating system and applications

NFS
WWW
E-mail
ssh
RPC
Transport Layer
IP
Network Layer
Physical/Link Layer
TCP
UDP
Ethernet
ATM
Packet radio
Network Protocols

Networking examples: many levels
– Physical level: mechanical and electrical network (e.g. how
are 0 and 1 represented)
– Link level: packet formats/error control (for instance,
the CSMA/CD protocol)
– Network level: network routing, addressing
– Transport Level: reliable message delivery

Layering – build complex services from simpler
services
Protocols
Physical Reality: packets
Abstraction: messages
Limited size
Arbitrary size
Unordered (sometimes)
Ordered
Unreliable
Reliable
Machine-to-machine
Process-to-process
Only on local area net
Routed anywhere
Asynchronous
Synchronous
Insecure
Secure

Our goal in the following is to show how to construct a
secure, ordered, message service routed to anywhere
Basic Networking Limitations

The physical/link layer is pretty limited
– Packets are of limited size (called the “Maximum
Transfer Unit or MTU: often 200-1500 bytes in size)
– Routing is limited to within a physical link (wire) or
perhaps through a switch
– Packets can get lost or garbled
– Physical routers crash/links get damaged
Datagram: an independent, self-contained network
message whose arrival, arrival time, and content are
not guaranteed
 Need resilient routing algorithms to send messages on
wide area

– Multi-hop routing mechanisms
– Redundant links/Ability to route around failed links
Arbitrary size messages


Arbitrary size messages on top of limited size ones
Send N little messages – split up message into fixed
size packets
– abcdefgh -> 1 of 3/abc 2 of 3/def 3 of 3/gh

Checksum can be computed on each fragment, or on
whole message
IP – Internet Protocol

a.
b.
Deliver messages unreliably (“Best effort”) from one
machine in internet to another.
Routes packets from one machine through internet
to another
Some intermediate links may have limited size.
–
c.

Fragments on demand, re-assembles at destination
Unreliable, unordered, machine->machine
Includes 256 different “sub-protocols” built on top
of IP
–
Examples: ICMP(1), TCP(6), UDP (17), IPSEC(50,51)
IP Packet Format
Flags &
Fragmentation
0
15 16
31 to split large
messages
4
IHL
ToS
Total length(16-bits)
16-bit identification flags 13-bit frag off
IP header
TTL
protocol 16-bit header checksum
20 bytes
32-bit source IP address
32-bit destination IP address
options (if any)
IP Header
Length
IP Ver4
Time to
Live (hops)
Type of
transport
protocol
Size of datagram
(header+data)
Data
Process-process communication
User process communication on top of machine to
machine communication
 What we really want is routing from processprocess

– Example: ssh, email, ftp, web browsing
Mailbox (or “port”) address – include in each message,
the destination mailbox.
 Allows you to direct each message to correct process
 A “port” is a 16-bit identifier used in addition to IP
address

– A communication channel (connection) defined by 4 items:
[source address, source port, dest address, dest port]
UDP – Unreliable Datagram Protocol
Unreliable, unordered, user-to-user communication
 Built on top of IP (IP Protocol 17)
 Important aspect: low overhead!

– Often used for high-bandwidth video streams
IP Header
(20 bytes)
16-bit source port
16-bit UDP length
16-bit destination port
16-bit UDP checksum
UDP Data
Ordered messages
Ordered messages on top of unordered ones
 IP can re-order packets – send A, B arrives: B, A
 How do we fix this? Assign sequence numbers to
successive packets –

– 0, 1, 2, 3, ...
If arrive out of order, don’t deliver #3 to user
application until get #2.
 Sequence numbers specific to a connection – for
example, the machine-machine (or mailbox-mailbox)
pair.
 This means put “source” as well as “destination” in
each header.

Performance considerations
Overhead – CPU time to put packet on wire
 Latency – how long before first bit of packet arrives
at receiver

– Raw transfer time + overhead at each routing hop

Throughput – maximum bytes per second
– Depends on “wire speed”, but also limited by slowest
router (routing delay) or by congestion at routers

Contributions to Latency
– Wire latency: depends on speed of light on wire
» about 1.5 ns/foot
– Router latency: depends on internals of router
» Could be < 1 ms (for a good router)
» Question: can router handle full wire throughput?
Example


How long to send 4KB packet over various networks? Typical overhead to
send a packet: 1 ms.
Ethernet (10 – 10,000Mb/s) within Parker:
– Latency: speed of light = 1.5 ns / foot, implies < 1 microsecond (if no
routers in path).
– Throughput delay: packet doesn’t arrive until all its bits get there! So
4KB/10 Mb/s = 3 milliseconds (roughly as long as a disk!)

ATM (155 Mb/s) within Parker:
– Latency same.
– Throughput delay: 4KB/ 155 Mb/s = 200 microseconds.

ATM cross-country?
– Latency: 3000 miles * 5000 ft/mile => 15 milliseconds.
– Throughput delay: same as above.

How many bits are in transit at the same time?
– 15 ms * 155 Mb/s => 280 KB

Key to good performance: in local area, minimize overhead, improve
bandwidth. In wide area, keep pipeline full.
Reliable message delivery


1.
2.
Reliable message delivery on top of unreliable
delivery
All of these networks can garble, drop messages.
Physical media – if transmit close to maximum rate,
get more throughput, even if some messages get lost
Congestion – what if no place to put incoming
message (no buffer space)?
–
–
–
–
Point-to-point network: no space at each switch/router
Broadcast network: two hosts try to use same link?
In any network: no buffer space at destination
Rate mismatch: what if sender sends faster than
receiver can process?
Reliable message delivery
1.
2.
3.
Detect garbling at receiver via checksum, discard if
incorrect
Receiver ack’s if received properly
Timeout at sender. If no ack, retransmit
Reliable message delivery


Some questions:
If the sender doesn’t get an ack, does that mean the receiver
didn’t get the original message?
– No.

What if ack gets dropped? Or if message gets delayed.
– Sender doesn’t get ack; retransmits. Receiver gets message
twice, ack’s each.

1.
2.
Solution: put sequence number in message to identify retransmitted packets. Receiver checks for duplicate sequence
#’s. If so, discards.
Sender must keep copy of every message that has not been
ack’ed yet (easy)
Receiver must keep track of every message that could be a
duplicate (hard! How does receiver know when it’s ok to forget
about received messages?)
Reliable message delivery

a.
Several approaches to maintaining state at sender/receiver:
Alternating bit protocol. One bit sequence number.
–
–
–
–

Pros:
–

Send one message at a time;
don’t send next message until ack received.
Sender only keeps copy of last message;
receiver keeps track of sequence # of last message received.
Simple, Small overhead
Cons
–
Poor performance
»
–
Wire can hold multiple messages; want to
fill up at (wire latency  throughput)
doesn’t work if network can delay
or duplicate messages arbitrarily
A
msg,#0
ack,#0
msg,#1
ack,#1
msg,#0
ack,#0
B
Reliable message delivery
b.
Window-based protocol (TCP).
–
–
–
–
Send up to N messages at a time, without waiting for ack.
“Window” reflects storage at receiver – sender shouldn’t
overrun receiver’s buffer space.
Each message has sequence number.
Receiver can say, “I’ve ack’ed up to message #X -> any
message below X is a duplicate.
B
A
msg,#0
ack,#0
Reliable message delivery

Acks serve dual purpose:
– Reliability: Confirming packet received
– Flow Control: Receiver ready for packet
» Remaining space in queue at receiver can be returned with
ACK

What if message gets garbled/dropped? Receiver will
get messages out of order!
– Discard any messages that arrive out of order?
» Simple, worse performance
– Keep copy until sender fills in the missing piece?
» Reduces # of retransmits, more complex

What if ack gets dropped?
– Timeout and resend just the un-acknowledged message.
TCP: transmission control protocol
Reliable byte stream between two processes on
different machines over Internet (read, write, flush).
 Fragments byte stream into packets, hands packets to
IP.
 Uses window-based protocol (to minimize state at
sender and receiver): send up to N messages at a
time, without waiting for ack.
 “Window” reflects storage at receiver – sender
shouldn’t overrun receiver’s buffer space.
 Sender has three regions: sent and ack’ed, sent and
not ack’ed, not yet sent

TCP: transmission control protocol

Receiver has three regions: received and ack’ed (given to application),
received and buffered, not yet received (or received and discarded
because out of order)
Sequence numbers
Sender messages
Sent,
ack’ed
Sent, not
ack’ed
Not yet
sent
receiver messages
Received,
given to
app

Received,
buffered
Not yet
received
Each ack says: “got all messages up to #”. What happens if ack is
delayed, arrives out of order? OK in this scheme. Just discard.
Arbitrary Size Messages (revisited)

1.
2.
3.
Face similar issues as in TCP when building big
messages on small ones, when messages can get
dropped.
Ack each fragment? Lots of acks
One ack for entire big message? Re-transmit all
fragments, even if only one gets dropped
“Blast protocol” – send one ack, tells sender which
pieces were missing. Selective retransmit.
Initialization

How do you know which sequence # to start with?
–
When machine boots, ok to start with #0?
»
»
–
1.
Thus, can re-use sequence numbers if wait for all packets in
flight to be delivered or to expire.
Epoch # – uniquely identifies which set of sequence numbers
are being used.
–
3.
Also, if it is possible to predict sequence numbers, might be
possible for attacker to hijack TCP connection
Time to live: each TCP packet has a deadline. If not delivered
in X seconds, then dropped.
–
2.
No. Could send two separate messages with the same serial #!
Receiver might end up discarding valid packets, or duplicate ack
from original transmission might hide lost packet
Put in every message, epoch # incremented on crash and/or
when run out of sequence #’s, and stored on disk.
Pseudo-random increment to previous sequence number
–
Used by a number of implementations now
Congestion

How long should timeout be for re-sending messages?
– Too long? Wastes time if message is dropped.
– Too short? Retransmit even though ack will arrive shortly.




Stability problem: more congestion -> ack is delayed ->
unnecessary timeout -> more traffic -> more congestion
TCP solution: “slow start”. Originally, window size = buffer space
on remote end.
Now, window size = control on how much to add to congestion.
Start sending slowly.
– If no timeout, slowly increase window size (throughput).
– If a timeout occurs, it means there’s congestion, so cut window
size (throughput) in half.
General’s Paradox
Can I use messages and retries over an unreliable
network to synchronize two machines so that they are
guaranteed to do some operation at the same time?
 Remarkably, no, even if all messages get through.
 General’s paradox: two generals, on separate
mountains.

– Can only communicate via messengers; the messengers can
be captured.

Need to coordinate the attack
– if they attack at different times, then they all die.
– If they attack at the same time, they win.
General’s Paradox
A
11 am ok!
B
Ok 11’s good for me
So 11 it is?
Yeah, but what if you
don’t get this ack?
...........

Even if all messages are delivered, can’t coordinate!
– Can’t simultaneously get two generals or two machines to
agree to do something at the same time.

No solution to this – one of the few things in CS that’s
just impossible.
Two phase commit
Since we can’t solve the General’s Paradox (i.e.,
simultaneous action), let’s solve a related problem.
 Abstraction: distributed transaction – two machines
agree to do something, or not do it, atomically (but
not necessarily at exactly the same time).

Two phase commit

Two phase commit protocol does this:
– Use a persistent, stable log on each machine to keep track of
whether commit has happened.
» If a machine crashes, when it wakes up it first checks its log to
see what state the world was in at the time of the crash.
– Prepare Phase:
» The global coordinator requests that all participants will promise to
commit or rollback the transaction
» Participants record promise in log, then acknowledge
» If anyone votes to abort, coordinator writes “abort” in its log and
tells everyone to abort; each records “abort” in log
– Commit Phase:
» After all participants respond that they are prepared, then the
coordinator writes “commit” to its log
» Then asks all nodes to commit; they respond with ack
» After receive acks, coordinator writes “got commit” to log
– Log can be used to complete this process such that all machines
either commit or don’t commit
Two phase commit
First phase, ask if each can commit – for instance,
transfer of funds from one bank to another.
1. A writes, “Begin transaction” to log
2. A -> B: OK to transfer funds to me?
 Not enough cash:
B-> A: transaction aborted
A writes “Abort” to log
 Enough cash:
B: Write new X account balance to log
B->A: OK, I can commit

Two phase commit


Second phase, A can decide for both, whether they will commit.
3. A: Write new Y account balance to log
4. Write commit to log
5. Send message to B that commit occurred
6. Write “Got commit” to log
What if:
– B crashes at 1? Wakes up, does nothing. A will timeout, abort
transaction, retry.
– A crashes at 3? Wakes up, sees transaction in progress. What
transaction, sends message to B, abort.
– B crashes at 3? B will come back up, look at log, so that when
A sends it “Commit” message, it will say, oh, ok, commit.
Two phase commit


One problem with 2PC is “Blocking”: That is, a site gets stuck in
a situation where it cannot continue until some other site
(usually the coordinator) recovers.
How could this happen?
– Participant site B writes a “prepared to commit” record to its
log, sends a “yes” vote to the coordinator (site A) and crashes.
– Site A crashes
– Site B wakes up, checks its log and realizes that it had voted
“yes” on the update.
» It sends a message to site A, asking what happened.
» At this point, B cannot change its mind and decide to abort,
because the update may have committed while it was crashed, thus
it is blocked.
» Note, B may be able to learn the fate of the update by asking
some of the other participants.
Two phase commit
Blocking is problematic because a blocked site must
hold resources (for example, locks on updated items,
pages pinned in memory, etc.) until it learns the fate
of the update.
 Question: Can blocking be avoided?
 Answer: yes! If you are willing to impose some
constraints on the way that voting is done, an
algorithm called “Three Phase Commit” can solve the
problem.

– 3PC is not generally used in practice, however, due to
performance reasons and the low instance of blocking in
practice.